No narrowing warnings when using `emplace_back` instead of `push_back` - c++

Question

My colleague ran into an unexpected issue with emplace_back and I am trying to wrap my head around it. The following test.cpp is a minimal example that reproduces the issue:
#include <vector>
class A {
public:
explicit A(int /*unused*/) {}
};
int main() {
double foo = 4.5;
std::vector<A> a_vec{};
a_vec.emplace_back(foo); // No warning with Wconversion
A a(foo); // Gives compiler warning with Wconversion as expected
}
Compiling with g++ 8.3.0 yields the following warning:
$ g++ -Wconversion test.cpp -o test
test.cpp: In function ‘int main()’:
test.cpp:10:10: warning: conversion from ‘double’ to ‘int’ may change value [-Wfloat-conversion]
A a(bar); // Gives compiler warning with Wconversion as expected
So the implicit conversion is caught when a simple object is constructed, but not when emplace_back is called.
Why is there no warning for emplace_back?

Answer 1

This is a consequence of how the default allocator constructs an A. When you do A a{foo, bar} you are using list initialization and a narrowing conversion is required to issue a diagnostic. With the default allocator, it uses
::new (static_cast<void*>(p)) T(std::forward<Args>(args)...)
where p is a pointer to the element of the vector data and T is the value_type of the vector. Here they use parentheses instead of braces and with parentheses narrowing conversions are allowed so you don't see a diagnostic message.
If you wrote your own allocator that did
::new (static_cast<void*>(p)) T{std::forward<Args>(args)...}
Then you would get the warning.

Answer 2

In order to get the warning you need to compile with:
$ g++ -Wsystem-headers -Wconversion test.cpp -o test
In file included from /usr/include/c++/8/vector:60,
from test.cpp:1:
/usr/include/c++/8/bits/stl_algobase.h: In function ‘constexpr int std::__lg(int)’:
/usr/include/c++/8/bits/stl_algobase.h:1001:44: warning: conversion from ‘long unsigned int’ to ‘int’ may change value [-Wconversion]
{ return sizeof(int) * __CHAR_BIT__ - 1 - __builtin_clz(__n); }
^
/usr/include/c++/8/bits/stl_algobase.h: In function ‘constexpr unsigned int std::__lg(unsigned int)’:
/usr/include/c++/8/bits/stl_algobase.h:1005:44: warning: conversion from ‘long unsigned int’ to ‘unsigned int’ may change value [-Wconversion]
{ return sizeof(int) * __CHAR_BIT__ - 1 - __builtin_clz(__n); }
^
In file included from /usr/include/x86_64-linux-gnu/c++/8/bits/c++allocator.h:33,
from /usr/include/c++/8/bits/allocator.h:46,
from /usr/include/c++/8/vector:61,
from test.cpp:1:
/usr/include/c++/8/ext/new_allocator.h: In instantiation of ‘void __gnu_cxx::new_allocator<_Tp>::construct(_Up*, _Args&& ...) [with _Up = A; _Args = {double&}; _Tp = A]’:
/usr/include/c++/8/bits/alloc_traits.h:475:4: required from ‘static void std::allocator_traits<std::allocator<_Tp1> >::construct(std::allocator_traits<std::allocator<_Tp1> >::allocator_type&, _Up*, _Args&& ...) [with _Up = A; _Args = {double&}; _Tp = A; std::allocator_traits<std::allocator<_Tp1> >::allocator_type = std::allocator<A>]’
/usr/include/c++/8/bits/vector.tcc:103:30: required from ‘void std::vector<_Tp, _Alloc>::emplace_back(_Args&& ...) [with _Args = {double&}; _Tp = A; _Alloc = std::allocator<A>]’
test.cpp:9:25: required from here
/usr/include/c++/8/ext/new_allocator.h:136:4: warning: conversion from ‘double’ to ‘int’ may change value [-Wfloat-conversion]
{ ::new((void *)__p) _Up(std::forward<_Args>(__args)...); }
The warning is no longer as clear, since it happens in a system header. Note that it is not enough to only use the -Wsystem-headers flag, one needs both -Wsystem-headers and -Wconversion to catch this.
In the case of push_back it is enough to use -Wnarrowing to get a warning.